DiSCo: Making Absence Visible in Intelligent Summarization Interfaces

TL;DR

This work tackles the problem of presence bias in AI-driven summaries by introducing DiSCo, a Domain-informed Summarization through Contrast framework. DiSCo constructs Domain Topical Expectations and computes deviations via LvS against each accommodation, surfacing both unusually frequent and missing domain-prevalent topics to enrich summaries generated by LLMs using structured prompts. Through a multi-domain user study (Ski, Beach, City-center), DiSCo improved perceived detail, domain relevance, and decision-support compared with baseline presence-only summaries, albeit with a modest increase in reading effort. The findings demonstrate that making absences explicit can enhance transparency and decision support, and they highlight the potential for adaptive, absence-aware interfaces across AI-assisted decision-making tasks.

Abstract

Intelligent interfaces increasingly use large language models to summarize user-generated content, yet these summaries emphasize what is mentioned while overlooking what is missing. This presence bias can mislead users who rely on summaries to make decisions. We present Domain Informed Summarization through Contrast (DiSCo), an expectation-based computational approach that makes absences visible by comparing each entity's content with domain topical expectations captured in reference distributions of aspects typically discussed in comparable accommodations. This comparison identifies aspects that are either unusually emphasized or missing relative to domain norms and integrates them into the generated text. In a user study across three accommodation domains, namely ski, beach, and city center, DiSCo summaries were rated as more detailed and useful for decision making than baseline large language model summaries, although slightly harder to read. The findings show that modeling expectations reduces presence bias and improves both transparency and decision support in intelligent summarization interfaces.

Figures (7)

• Figure 1: DiSCo pipeline. The process of Domain-informed Summarization through Contrast (DiSCo) integrates expectation-based analysis with LLM-based summarization. (1) User reviews are collected for all accommodations in a predefined area. (2) Aspect and sentiment extraction identifies topics and their polarity. (3) LvS computes divergences between each accommodation’s topic distribution and domain-level topical expectations, revealing over-represented and missing (expected but unmentioned) aspects. (4) These deviations are incorporated into a structured LLM prompt for summary generation. (5) The resulting DiSCo summaries explicitly surface both salient and absent information relative to domain norms.
• Figure 2: LvS-based analysis framework for identifying accommodation-specific topic deviations from reference distribution mokryn2025interpretable. (A) Input data: topic distributions within individual accommodations, where colored circles represent distinct topics and circle size indicates topic prevalence. These were identified in phase 2 of DiSCo. (B) Aggregate reference distribution computed across all accommodations in the corpus. (C--D) Accommodation-specific deviation analysis showing topics appearing in surplus (blue bars, above reference baseline) or absent (red bars, below reference baseline). Blue annotations indicate topics exceeding expected frequency, and red annotations indicate popular topics that are underrepresented or absent in the focal accommodation.
• Figure 4: Example of presence–absence topic analysis for a single accommodation. Topics that appear more frequently than the domain average are shown on the right, while domain popular topics that are missing or hardly mentioned are shown on the left. Positive values indicate over-representation, and negative values indicate under-representation or absence.
• Figure 5: Example of a study presenting two versions of a summary for a single accommodation
• Figure : (a) City-center domain